Current H100 cloud pricing
Today’s sourced per-provider rates and the CT-H100 read.
← Back to Compute College
Compute College
NVIDIA H100 Price Per Hour Explained
H100 price per hour is the hourly cost of accessing one NVIDIA H100 GPU for an AI workload.
H100 price per hour definition
NVIDIA H100 price per hour is the hourly cost to rent or operate one H100 GPU for AI workloads. It is usually expressed as an H100 GPU-hour rate, but real buyer cost depends on provider, contract type, available cluster size, networking, support, region, and utilization.
Memory trick: Think of the H100 price as a yardstick: useful for measuring the market, but not a complete description of every workload.
Live price band
H100 on-demand price band
H100 on-demand capacity ranges roughly $3.29–$12.29 per GPU-hour across 5 sourced providers, as of Jul 7, 2026. Each row below links to the provider's public price page and carries its own observation date.
Public on-demand list prices normalized to a per-GPU-hour rate — not negotiated quotes or reserved pricing. How we label and date evidence: methodology.
Sourced providers
Per-provider H100 on-demand rates
| Provider | Region | $/GPU-hour | Source | Observed |
|---|---|---|---|---|
| RunPod Secure | Multi-region | $3.29 | price page | Jul 7, 2026 |
| Crusoe | Multi-region | $3.90 | price page | Jul 7, 2026 |
| Lambda | Multi-region | $4.29 | price page | Jul 7, 2026 |
| AWS | us-east-1 | $6.88 | price page | Jun 21, 2026 |
| Azure | eastus | $12.29 | price page | Jun 21, 2026 |
Why it matters
Why H100 hourly pricing matters
The H100 is a useful reference point for modern AI compute pricing because buyers frequently compare offers against H100 capacity even when considering H200, B200, or another accelerator. A clear H100 benchmark helps make quote changes, availability, and hardware transitions legible.
- A familiar baseline helps buyers compare providers and contract structures.
- H100 availability can matter as much as rate when a workload needs a cluster on a deadline.
- Price movement can show how demand shifts as newer accelerator generations become available.
Simple example
Assume, only for comparison, that a provider quotes $7.00 per H100-hour. Eight H100 GPUs for 10 hours would cost 8 x 10 x $7.00 = $560 before overhead. Another offer at $5.00 per hour is not automatically better if it is interruptible, unavailable at the required cluster size, or missing needed network services.
- Normalize every quote into the same H100 GPU-hour quantity.
- State whether the offer is on-demand, reserved, or spot capacity.
- Then compare workload completion, reliability, and extra service costs.
Example figures are illustrative calculations, not current quoted market prices.
Market signal
How to read the market signal
A rising comparable H100 rate can indicate tighter supply, stronger short-term demand, less discounted capacity, or buyers paying for dependable access. A falling rate can reflect additional supply, provider competition, lower demand, or workloads migrating toward newer hardware.
- Check whether rates move together across providers before treating one quote as a market move.
- Watch availability and reservation terms: scarcity may appear there before the hourly number changes.
- Compare H100 rates with H200 and B200 premiums to read the hardware transition.
Market read: H100 is a benchmark reference, not a single universal price. ComputeTape publishes current sourced H100 cloud pricing, the CT-H100 median, and per-provider public list rates on the H100 pricing page before any single quote is treated as representative. A useful market read compares like-for-like capacity, availability, and terms across more than one provider observation. Figures here are illustrative unless explicitly sourced and dated — see our methodology.
Common mistake
Do not compare headline H100 prices without checking the product underneath them. A single GPU, a tightly connected multi-GPU node, bare metal, a virtual machine, reserved capacity, and an interruptible offer all deliver different economic value even when each advertises an H100-hour.
Practical takeaway
What you can do with this
Use H100 hourly pricing as a benchmark reference, then require comparable terms. Buyers should record cluster size, contract duration, interruption risk, networking, support, region, and utilization assumptions next to the quoted rate.
- Procurement teams: compare effective completed-workload cost, not just price per rented hour.
- Founders: budget both the quoted capacity and the risk of delayed access.
- Analysts: separate broad benchmark movement from a provider-specific promotion or constraint.
- Buyers: ask whether a quoted rate remains available at the required start time and cluster size.
- Teams comparing alternatives: document why any H100 premium is acceptable before signing a commitment.
Decision check: reject comparisons that show an hourly rate without the matching availability, contract, cluster, and workload-performance conditions.
Compute College
Turn the lesson into a number
Use the GPU-Hour Cost Calculator, AI Training Cost Calculator, or Model Serving Cost Calculator.
Compute College track
GPU Pricing & Capacity
Step 1 of 8: H100 price per hour